Large Language Model probabilities cannot distinguish between possible and impossible language